Selected issues of semantic analysis 1. Normalization and paraphrasing

Semantic analysis rules operate on Normalized Syntactic Structure and produce Basic Semantic Structures (SemS, see Section 3 below). One of the first tasks that should be done is the canonization. It includes restoring subjects of non-finite verbs (I want to run → I want: I run), processing of ellipsis, comparative constructions and the like. We will illustrate one canonization pattern: elimination of semantically void collocates. This operation is performed by means of a paraphrase generator based on Lexical Functions (LF) (Mel’čuk 1996; Apresjan, Cinman 2002). The paraphrase generator is a system of rules relying on a rich dictionary of lexical functions. In the semantic analyzer, the generator reduces sentences containing collocate LFs to the canonical form without these LFs. Some examples are:

(4) → (5), (6) → (7), (8) → (9).

(4) John has respect for his teachers / John’s teachers enjoy his respect / John treats his teachers with respect

(5) John respects his teachers

(6) The police gave the protesters an order to disperse / The protesters were ordered by the police to disperse / The protesters received an order from the police to disperse.

Semantic representation for NL understanding

(7) The police ordered the protesters to disperse.

(8) The experts should submit / prepare / make / produce a report on chemical weapons.

(9) The experts should report on chemical weapons.

Strictly speaking, the sentences in these pairs are not fully synonymous. How- ever, the semantic representation we are striving at does not aim to account for all subtleties of meaning. The level of granularity of semantic representations should be determined by the task for which they are constructed. The immediate objective of our semantic representations is to support inference. For this aim, semantic differ- ences that can be observed in these pairs are not relevant. Should an application re- quire finer-grained representations, paraphrasing rules should be made more precise.

An example of a subtler representation is given in the next section.

2.2. Semantic definitions of NL words and ontology concepts

Deep NL understanding requires much more elaborated meaning representa- tion than that provided by standard resources. For example, semantic parsers based on FrameNet annotate verbs like ask for, implore and demand by relating them to the same REQUEST frame. It is true that all the three verbs have the same set of roles. The generalized frame REQUEST allows us to capture this similarity, but we also want to preserve the knowledge about their difference. First, one should make explicit the very idea of the incentive speech act (e.g. ‘X tells Y that he wants him to do Z’). This should be made in a formal language so that it could be used for inferences. Second, since our ultimate aim is to model natural language as fully as possible, it is desir- able to account for the semantic difference between the varieties of this speech act.

There exist many NL speech act types in which the agent informs the addressee that he wants the latter to do something.

Below, we give definitions of three of them: ask (as in He asked to open the window), implore (as in They implored her to help) and demand (as in She demanded an explanation). Roughly, the difference between ask and demand is that the one who is asking does not think that the addressee is obliged to fulfill the request, while the one who is demanding assumes that the addressee must do it. Imploring adds to ask- ing the idea that fulfilling the request is very important for the agent so in persuading the addressee to do it he tries to affect his feelings. In the definitions, variables are marked with the ? sign. For brevity, the ontological class to which the variable belongs is encoded by the name of the variable.

(10) ask for (?Agent1,?Agent2,?Action) [=‘?Agent1 tells ?Agent2 that he wants him to do ?Action; ?Agent1 does not think that ?Agent2 must do ?Action’]

hasAgent(Tell,?Agent1) hasRecipient(Tell,?Agent2)

Boguslavsky I. M., Dikonov V. G., Iomdin L. L., Timoshenko S. P.

hasObject(Tell,Want) hasSubject(Want,?Agent1) hasObject(Want,?Action) hasAgent(?Action,?Agent2) hasScope(Negation,Opinion) hasSubject(Opinion,?Agent1) hasObject(Opinion,?Action) hasScope(MustModality,?Action)

(11) implore (?Agent1,?Agent2,?Action) [=‘?Agent1 asks ?Agent2 to do ?Action;

it is very important for ?Agent1 that ?Agent2 realizes ?Action; ?Agent1 tries to affect the feelings of ?Agent2’]

SemS of (11) consists of the SemS of (10) plus the following:

hasSubject(Important,?Action) hasObject(Important,?Agent1) hasSubject(Degree,Important) hasValue(Degree,high)

hasAgent(Affect,?Agent1) hasObject(Affect,Feeling) hasSubject(Feeling,?Agent2)

(12) demand (?Agent1,?Agent2,?Action) [=‘?Agent1 tells ?Agent2 that he wants him to do ?Action; ?Agent1 thinks that ?Agent2 must do ?Action’]

hasAgent(Tell,?Agent1) hasRecipient(Tell,?Agent2) hasObject(Tell,Want) hasSubject(Want,?Agent1) hasObject(Want,?Action) hasAgent(?Action,?Agent2) hasSubject(Opinion,?Agent1) hasObject(Opinion,?Action) hasScope(MustModality,?Action) hasAgent(?Action,?Agent2)

2.3. Converse terms

Natural languages have hundreds of converse terms, i.e. pairs of words that de- note the same situation but differ in the syntactic status of their arguments. Obvious examples are husband — wife, buy — sell, to the right of — to the left of, more — less, better — worse, etc. Although these words are not synonyms, if we swap positions of arguments we obtain equivalent assertions:

(13) John is Mary’s husband = Mary is John’s wife.

Semantic representation for NL understanding

(14) John bought a house from Mary = Mary sold a house to John.

(15) The table is to the right of the window = The window is to the left of the table.

(16) John likes physics more than geography = John likes geography less than physics.

Since converse terms refer to the same situation, it is sufficient for a semantic lan- guage and ontologies to contain only one term of the pair. In our semantic language, we have only one correlate for the ‘more’/‘less’ pair — concept MORE. In represent- ing this meaning, we differ from some other approaches (such as e.g. (Nirenburg, Raskin 2004), which treat ‘more’ as a binary relation: A>B. Our MORE concept has three arguments: A — “what is more?”, B — “more than what?”, C — “by how much is A more than B?”. In sentence (17) the arguments of MORE are: A=John’s height, B=Bill’s height, C=3 cm.

(17) John is 3 cm taller than Bill.

On the other hand, one can opt for having both members of the converse pair. For in- stance, we represent husband and wife by different concepts, because these social roles are bound by different conventions and stereotypes which have to be described in the ontology.

An interesting case of converse relations, which as far as we are aware was first introduced in (Boguslavsky 2009), is the relationship between all and only.

(18) Here are all my documents

(“for any document x of mine, it is true that x is here”).

(19) Here are only my documents

(“for any x that is here, it is true that x is my document”).

This allows us to have only one semantic unit — a two-place predicate All, which covers both all and only. Here are semantic structures for sentences (22) and (23):

(20) All the children who guessed the riddle got a prize.

(21) Only the children who guessed the riddle got a prize.

(20a) hasElements(Set,Child) hasAgent(Guess,Child) hasObject(Guess,Riddle) hasAgent(Get,Set) hasObject(Get,Prize) hasSubject(All,Set) hasObject(All,Get)

Boguslavsky I. M., Dikonov V. G., Iomdin L. L., Timoshenko S. P.

(21a) hasElements(Set,Child) hasAgent(Guess,Child) hasObject(Guess,Riddle) hasAgent(Get,Set) hasObject(Get,Prize) hasSubject(All,Get) hasObject(All,Set)

2.4. Evaluation of objects

Evaluation of objects and events plays an enormous role in our life, everyday behavior and common sense reasoning. Therefore the world knowledge modeled by the ontology should contain manifold information on what is good and, bad and for whom. For many situations, we are aware that they are either beneficial or detrimen- tal to the interests of some of their participants. For example, if somebody dies, is sick, late for an appointment, gets ruined, receives a rebuke, or fails an exam, by default this is bad for him. If, on the other hand, he recovers from an illness, gets an award, is promoted or attains his aim, then, again by default, it is beneficial for him. Some sit- uations are estimated differently from the point of view of their different participants.

For example, a victory (in a conflict, debate, sports competition, etc.) is beneficial for the winner and adverse for the loser. We will demonstrate that this kind of informa- tion can play a role in text understanding. Then we will show how it is incorporated in our Ontology and used for semantic analysis.

Consider sentence (22) and its two possible continuations — (23) and (24).

(22) In the first tour FC Spartak overwhelmed FC Dynamo.

(23) In the second tour FC Zenith suffered the same fate.

(24) In the second tour FC Zenith managed to achieve the same thing.

Both (23) and (24) contain the anaphoric expression the same that refers to sen- tence (22). In both cases, a situation is described that is similar to (22), the only differ- ence being that one of the clubs is replaced with Zenith. In (23) an analogy is drawn between Zenith and Dynamo, and in (24) between Zenith and Spartak. In other words, (23) is unambiguously understood as ‘Spartak overwhelmed Zenith’, while (24) means that ‘Zenith overwhelmed Dynamo’. It is noteworthy that even though nei- ther (23) nor (24) explicitly specifies the opponent of Zenith, it is “calculated” from the evaluation semantics.

To be able to draw these conclusions, the system should dispose of the following knowledge:

(a) “P is fate suffered by X” implies that P is not in the interests of X;

(a) “X managed to achieve P” implies that P was among X’s aims and P is benefi- cial for X;

Semantic representation for NL understanding

(a) “victory of X over Y” is beneficial for X but not for Y.

This knowledge is incorporated into the system as follows:

• The Ontology contains an Evaluation concept, which has 4 slots: the agent of the evaluation (hasAgent), the object or event under evaluation (hasOb- ject), the value of the evaluation (hasValue) — good or bad and the ben- eficiary, i.e. someone for whom the object or event is beneficial or adverse (hasBeneficiary).

• This concept is introduced into the description of the concepts which include a de- fault evaluation (cf. examples above). The WinEvent concept, which has slots for the winner (hasWinner) and for the loser (hasLoser) and which covers both a victory and a defeat, is assigned the following properties, among others:

hasWinner(WinEvent,?SportAgent1) hasLoser(WinEvent,?SportAgent2) hasObject(Evaluation-01,WinEvent) hasValue(Evaluation-01,good)

hasExperiencer(Evaluation-01,?SportAgent1) hasObject(Evaluation-02,WinEvent)

hasValue(Evaluation-02,bad)

hasExperiencer(Evaluation-02,?SportAgent2)

• A reference to evaluation is included into semantic rules that interpret natural language evaluating expressions. X suffered the fate of P contains the compo- nent “P is estimated to be bad for X”. In our semantic language it is represented as follows:

hasObject(Evaluation,P) hasValue(Evaluation,bad) hasBeneficiary(Evaluation,X)

• Expressions like X succeeded in / achieved P include in their definition a reference to P being the aim of X, which in its turn implies that P is beneficial for X:

hasObject(Evaluation,P) hasValue(Evaluation,good) hasBeneficiary(Evaluation,X)

Now, let us see how this knowledge helps interpret sentences (23) and (24).

As mentioned above, proposition (22) serves as the antecedent of ‘the same’, so theoret- ically, it can be introduced into the SemS of both (23) and (24) in two different ways:

(24a) hasWinner(WinEvent,Zenith) hasLoser(WinEvent,Dynamo)

(meaning that Zenith beat Dynamo like Spartak beat Dynamo) or (23a) hasWinner(WinEvent,Spartak)

hasLoser(WinEvent,Zenith)

(meaning that Zenith lost to Spartak like Dynamo lost to Spartak).

Boguslavsky I. M., Dikonov V. G., Iomdin L. L., Timoshenko S. P.

However, taking into account that the meaning of Zenith suffers a fate assigns to Zenith the role of the beneficiary of a negative evaluation, while in the WinEvent it is the winner who benefits, version (24a) should be rejected for sentence (23).

In a similar way, (23a) is rejected for (24).